JPS63234300A - Voice recognition - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a voice recognition method that allows a machine to recognize a human voice.

2. Description of the Related Art Speech recognition technologies have been actively developed and commercialized in recent years, but most of these are for specific speakers whose voices are recognized only by those who have registered their voices.

Devices for specific speakers require time and effort to register the words to be recognized in the device in advance, which puts a heavy burden on the user unless the device is used continuously for a long time. In response to this, research has recently been actively conducted on recognition technology for non-specific speakers that is easy to use and does not require voice registration.

Generally speaking, the speech recognition method performs pattern matching between the input speech and standard speech stored in a dictionary (these are parameterized), and selects the speech in the dictionary with the highest degree of similarity. This means that it is output as a recognition result. In this case, there is no problem if the input voice and the voice in the dictionary are physically exactly the same, but in general, even if the input voice is the same, different people say it or say it in different ways, so they may not be exactly the same. It won't be.

Physically, differences between people and differences in the way they speak are expressed as differences in spectral features and differences in temporal features.

In other words, the shape of the articulatory organs (mouth, tongue, throat, etc.) differs from person to person, so the spectral shape of the same word will differ between different people.

Furthermore, the temporal characteristics differ depending on whether the voice is spoken quickly or slowly.

Speaker-independent recognition techniques require such spectra and their temporal variations to be normalized and compared to standard patterns.

As an effective method for speaker-independent speech recognition, the present applicant and others have already proposed a method that uses parameter time-series information and a statistical distance measure (Niyata et al.: “Simple speaker-independent speech recognition”). "Speech Recognition Method for Use", Proceedings of the Acoustical Society of Japan, 1-
1-4 (March 1986)), the method will be explained below.

This method uses a pattern matching method to spot speech in noise, thereby making it possible to recognize speech and detect speech sections at the same time.

The distance measure (statistical distance measure) used in mass and pattern matching will be explained.

If the input word audio length is linearly expanded or contracted to J frames and the parameter vector per frame is set to XI, the input vector X becomes as follows.

X = (machi, taka 2..., scent) Here, each aJ is a one-dimensional vector.

As a standard pattern for the word ωh (k=1. 2...,), let the mean value vector be μ and the covariance matrix 1 be lWk
Then, the word with the maximum posterior probability P(ωk 1st) may be taken as the recognition result.

From Bayes theorem, P(ωkl &) = P(ωk) - P(Xlωk)/P(X
) (1) P(ωk), the first term on the right-hand side, can be regarded as a constant. Assuming a normal distribution, the second term is P(&1cuk)=(2ffE"1lWkl-"・ex
p(-1/2(x-Iuk)-1W k-(J-tuk
>) (2) The denominator term P(X) can be regarded as a constant if the input parameters are the same, but it does not become a constant when comparing different inputs. Here, it is assumed that P(&) follows a normal distribution with mean value/Ijk and covariance matrix Wx.

p (x) = (2yr) 21! vl/x
l 2 · axp (-1/2 (J-ax) WV' - (, X-ruk)) (3) If we take the logarithm of (1), omit the constant term, and set it as lLk, rLk=(X-11h)-1pt-'h-(X-tuk
)-(X-1uk) ・'fl 'x ・(J(-μ
k)+logl"l/l/kl-1ogl'Wkl
(4) Here, in ll#, set v/x to be common to all and set it to \W.

That is, sw=(w,-)! /2+-・--+g+nt-+w
x)/(K+1) (5) Expanding formula (43), Lk=Bk-^ becomes -X (
a) However, 1Ast=2(vv--tuk-w-#k)
(7) Bk = Iuk ”W-・Red-fu
k QW-' 4k (8) Equation (6) is a first-order discriminant with a small amount of calculation. Here, (@expression is transformed as follows.

Assuming that Ah = (pray for one gourd), ..., ni)), that is, Lk is the partial similarity for each frame, (k, l =
, ts=3 ・Machi is calculated by - times of addition and once of subtraction.

Next, a method for spotting and recognizing speech in noise using the above distance measure and a method for reducing the amount of calculation will be explained.

Targeting a sufficiently long interval that definitely includes speech, set various subintervals within this, calculate the similarity with each word using the (@ formula, and find the word with the maximum similarity across all subintervals. This can be the recognition result.If this similarity calculation is performed as it is, the amount of calculation will be enormous, but by considering the duration of the word and limiting the partial ward island length, the partial similarity d( 7) can significantly reduce the amount of calculation. FIG. 3 is an explanatory diagram of this method.

When matching input to words, subinterval length n (n'
:kn<n'P)) to the standard pattern length J,
It shows how the similarity is calculated for each frame with a fixed end. The degree of similarity is calculated using equation (9) along a route starting from point T on QR and ending at P. therefore,
All similarity calculations per frame are performed within ΔPQR. By the way, the town in equation (9) is the 'th frame component after expanding and contracting the section length n, so the corresponding input frame 8'
exists.

Therefore, using the input vector, 1. (?) can be expressed as follows.

d"J+', ,)=j'7-a, (
1o) However, l'=藝−r(J)−H(11)Here, Pn(J) is the word length n,! : It is a function that relates the linear expansion and contraction of J. So each frame of input and &(
If the degree of partial similarity of 艷 is determined in advance, equation (9) can be easily calculated by selecting and adding the degree of partial similarity having the relationship of 良'. By the way, △PQR moves to the right every frame, so calculate the partial similarity of 1μ on the PS, store it in memory by the amount equivalent to ΔPQS, and shift it every frame. If configured as follows, all the necessary similarities are stored in the memory, so the operation for determining partial similarities can be largely omitted, and the amount of calculations can be extremely reduced.

FIG. 4 is a functional block diagram illustrating a conventional implementation method. The unknown input audio signal is converted into 8KH by the An converter 10.
The signal is z-sampled and converted into a 12-bit digital signal. The acoustic analysis unit 11 performs LPG analysis of the input signal every 10 m$·o (1 frame), and obtains the 10th-order linear prediction coefficient and residual power. The feature parameter extraction unit 12 uses the linear prediction coefficients and residual power to obtain the LPG cepstrum coefficients C1 to C5 and the power term Co as feature parameters. Therefore, the feature vector definition for each frame is A' = (C□ + 01, ...c s )
(12). In addition, LPG analysis and LP
Regarding the extraction method of C cepstral coefficients, for example, J, D
Since it is described in detail in "Linear Prediction of Speech" by Markel, AH, Gray, translated by Hisaki Suzuki, it will be omitted here.

The frame synchronization signal generator 13 is a part that generates a timing signal (frame signal) every 10 m$·C, and recognition processing is performed in synchronization with the frame signal.

The standard pattern selection unit 18 selects the word number stored in the standard pattern storage unit 17 during one frame period.
1.2. ...select one after another. The partial similarity calculation unit 21 calculates the partial similarity ti(k' (
Calculate re J ).

d(k) (river) Hiro(-hei(J=1.2...J)(1
3) The calculated partial similarity is output to the similarity buffer 22 and accumulated therein. The similarity buffer 22 is configured such that when a new input is input, the oldest information disappears.

For each selected word number, the section candidate setting unit 15
Set the minimum length, (k) and maximum length, (k) of the word.
e. The time expansion/contraction table 24 stores the relationship of equation (11) in a table format, and the word length n and frame J.
When , the corresponding l' is found. , (k)≦
$ for each word length n in the range n≦n(k)
・ Read 1 for the corresponding partial similarity d(k'
(1′, J), J = 1t 21 . . . J is read from the similarity buffer 22. The similarity adder 23 is inserted. The similarity comparison unit 20 temporarily stores 19
, and the one with greater similarity (smaller distance) is recorded in the temporary memory 19.

In this way, starting from frame l''Io, we have a standard pattern with n(-) <, <, (S) for =1
As a range, L calculated in the range of (2)<n<, (2)
Find the maximum similarity by comparing 2゜ and -. ' is stored in the temporary memory 19.Next, the same procedure is repeated with 1=IO+Δ1, and when the final frame l=1 is reached, the word number left in the temporary memory is =krn.The recognition result is also , if the frame number i=- and the word length n=nrn when the maximum similarity was obtained are stored in the temporary memory 19 and updated, the speech section at that time can be obtained as a result at the same time as the recognition result. The voice interval is from 1□-'m to -□.

Problems to be Solved by the Invention The problem with this method is that pattern matching is performed with all possible speech sections within a sufficiently long segment that definitely contains speech, so for example, when it comes to numeric speech, In the recognition of
There is a high possibility that the part of a long uttered word will be recognized as a short word, such as the chi part of ``chi'' being recognized as ``kyu'' or the ``i'' part of ``two'' being recognized as ``two''.

An object of the present invention is to solve the above-mentioned problems, and to solve the above problem, when each speech interval corresponding to the maximum similarity or minimum distance with respect to each standard pattern is abnormally different, the maximum similarity or minimum distance of each standard pattern is The purpose of this invention is to provide a speech recognition method that has a high recognition rate by correcting.

Means for Solving the Problems The present invention achieves the above object by calculating the maximum (minimum) similarity (distance) and the maximum similarity (minimum distance ti1m) between the feature parameter and the standard pattern. Find the voice that corresponds to the voice, add the values of the pre-prepared bias table corresponding to that voice to the similarity (distance) of the above, and calculate the standard that has the maximum (minimum) similarity (distance) among them. The speech corresponding to the pattern is the recognition result, and the maximum similarity corresponding to the standard/4 turns is defined as Lk, (1
≦-≦), then L 1 = n'llll x (Ll , L2 , "・-L
K ) ``・"・・"・" (14), and use a bias table prepared in advance consisting of x numbers as B (amount, ") (1≦薯≦).

1≦j≦K), then L' hm2max (L 1 +8 (1, 1)
, L2+8 (1,2).

..., LK-□ (to 1) ... (15), kr
n is the recognition result.

Function The present invention corrects the maximum similarity value of two standard patterns in a recognition method using a word spotting method that does not clearly define speech intervals for unspecified speakers. It is possible to reduce erroneous recognition and improve the overall recognition rate when each speech section corresponding to the maximum similarity value matched with each standard pattern is abnormally different.

Example Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a functional block diagram showing an implementation of a speech recognition method according to an embodiment of the present invention.

First, the basic recognition concept of this embodiment is almost the same as the method described in the conventional example. That is, the unknown input audio signal is sampled by Hχ at 8 in the An converter 110,
It is converted into a 12-bit digital signal. The acoustic analysis unit 111 performs LPG analysis of the input signal every 10 m5·O (one frame), and obtains the 10th-order linear prediction coefficient and residual power. The feature parameter extraction unit 112 uses the linear prediction coefficients and residual power to calculate the LPC cepstral coefficient &
C, ~C9 and the power term co are determined as feature parameters. Therefore, the feature vector for each frame is t = (CQ, c, ......C9
) (16). In addition, LPG analysis and L
Regarding the extraction method of PG cepstral coefficients, for example, J,
l) is described in detail in "The Linear Side of Speech" by Markel, A., H. Gray, translated by Hisakure Suzuki, so it will be omitted here.

The frame synchronization signal generation unit 113 is a part that generates a timing signal (frame signal) every 10 m@·C, and recognition processing is performed in synchronization with the frame signal.

The standard pattern selection unit 116 successively selects word numbers =1.2...K stored in the standard pattern storage unit 115 during one frame period. The calculation unit 114 calculates the selected standard pattern n(
k) and the partial similarity of the feature vector town of the first frame d(
k) Calculate (1,j).

-) (river) = α (1) 1・accent, (J=1.2...
J) (17) The calculated partial similarity is stored in similarity buffer 1.
19 and accumulates it. The similarity buffer 119 is configured such that when a new input is input, the oldest information disappears.

For each selected word number, the 8-word candidate setting unit 117 determines the minimum length n(k) and maximum length n(k) of the word.
・Set. The time expansion/contraction table 118 stores the relationship of equation (11) in a table format, including word length, (n(
k)≦n≦n(k)) $
6 and the corresponding white (find.n(k)≦n≦n(k
) for each word length n in the range of
．． 2...J is read from the similarity buffer 119. The robustness addition unit 120 calculates 嗟−k(,′・1),
Calculate the number of units using equation (9). Similarity comparison section 1
21 compares the obtained data with the contents of the temporary memory 122, and stores the one with a greater degree of similarity (smaller distance) in the temporary memory 122, and at the same time stores the one with the contents of the temporary memory 122 The maximum similarity Lkr of the stored similarities of
Find n and memorize it.

Start this kind of operation from frame l = 1° and move to 7L/
- When the purest frame I=1 is reached,
The similarity comparison unit 121 calculates the calculated maximum similarity krn
The value of krn, (i.e., 'm represents the krnth standard pattern that produced the maximum similarity), is determined by bias and
Send table 128'#ζ.

Bias table 128 has 8 (1, J)
It consists of the predetermined numerical value of , and when k is input, (-B(kfTl・J) (J=1・),
It is sent to the similarity comparison unit 121. The similarity comparison unit 121 calculates the maximum similarity ma for each standard pattern in the temporary storage 122.
Ask xL1k to be sent to you and calculate the following equation.

Yk'm = msx(maxLl'+8(km, 1)
.

maxL22+B(krn, 2).

...1m ron+a (k,,,,ni))...
(18) That is, among the terms on the right side of the equation <18), - is determined for the maximum term, and k- is set as the recognition result.

Next, a method of determining the scanning section from 1° to 1 in the above description will be explained.

FIG. 2 shows the start of scanning (start after cheek similarity addition unit 120)
This shows the relationship between the I0 frame, one frame of recognition completion, and audio.

The starting point of the scanning section is found using power information, and the ending point is found using both power information and similarity information. FIG. 1 (The D power calculation unit 123 calculates the average power (logarithm value) for each frame p+(
tit frame number) and sends it to the power comparison section 125. The power comparison section uses the output PN from the noise level learning section to compare Pl and PN+θN (θ, threshold level), and sets it to 1 if P is large and 0 otherwise. Output to the scanning section setting section. Also, PIkuPNNO
6, P is sent to the noise level learning section 124, and the noise level learning section uses P to recalculate PN. The calculation formula is: PN =, (FALL + PI)
('') where OPEN is the number of powers sent, PALL
is the sum of the power sent up to the previous frame.

shall be.

In the scanning section setting section, the first l is determined by the power comparison section 125.
When is output, scanning is started, and the similarity adder 120 and subsequent parts start operating. Further, the conditions for determining the end of scanning are as follows: 1) θ frames have passed since the last time the output of the power comparison unit 125 changed from 0 to 1, and a 0 frame exists.

2) 02 frames can be read from the time of the last change from 1 to O, and there is a 0 frame.

3) The maximum similarity 'km of the similarity comparison section is greater than θ3.

When the above three conditions are satisfied, recognition is terminated.

In the method described in the conventional example, in which the maximum similarity is calculated from all possible speech sections around a place that seems to be speech without determining the speech section, power information is generally used to determine the speech section. It can be said that this method is stronger than the method of determining and matching a standard pattern when the three-tone level is higher or when non-stationary noise is mixed, but on the other hand, if the recognition target word contains the l part of a long word and very If there are short words similar to , the recognition rate will be very poor.

For example, there is a case where "Shin-Osaka" and "Osaka" are included in the words to be recognized. In the case of this example, by utilizing the property that "Shin-Osaka" is likely to become "Osaka" but "Osaka" is difficult to become "Shin-Osaka", a bias table prepared in advance is applied to the similarity or distance obtained once. By adding the numerical values of , and using the maximum or minimum of them as the recognition result, partial matching can be prevented as much as possible, which is a very effective means.

Effects of the Invention In short, the present invention calculates the voice with the maximum or minimum similarity or distance, adds the numerical value of a bias table prepared in advance corresponding to that voice to the similarity or distance, and calculates the similarity or distance. The recognition result is the voice corresponding to the standard pattern with the maximum or minimum degree or distance.
This method has the advantage of preventing the recognition of words with similar utterances and improving the overall recognition rate.

[Brief explanation of the drawing]

Fig. 1 is a &n block diagram embodying the speech recognition method in one embodiment of the present invention, Fig. 2 is a relation diagram between the start and end times of standard pattern stop matching and speech in this embodiment, and Fig. 3 4 is a conceptual diagram illustrating a pattern matching method with a standard pattern, and FIG. 4 is a functional block diagram illustrating a conventional method. 110...AO conversion unit, 111...Acoustic analysis unit, 112...Feature parameter extraction unit,
It;L... Frame synchronization signal generation section, 114.
...Partial similarity calculation unit, 115...Standard pattern storage unit, 116...Standard pattern selection unit, 117...Section candidate setting unit, 118...・
...Time expansion/contraction table, 119...Similarity buffer, 120...Similarity addition unit, 121...
...Similarity comparison section, 122 ...Temporary memory,
123...Power calculation section, 124...
Noise level learning unit, 125...Power comparison unit, 127...Scanning section setting unit, 128...
...bias table. Part 2 Figure 3

Claims (4)

[Claims] (1) Detect the presence of voice using power information from an unknown input signal including voice and noise before and after the voice, and use the time of detection as a reference point, and from the reference point and reference point N (N_1≦N≦N
Linearly expand or contract the unknown input signal in a section separated by _2) to the section length L, extract the feature parameter of the expanded or contracted section, and calculate the similarity or distance between this feature parameter and the standard pattern of K voices to be recognized. , and then perform these operations while changing N from N_1 to N_2.Furthermore, perform similar operations while shifting the reference point by unit interval to successively obtain and compare the degrees of similarity or distance. Then, all reference points and all time expansion/contraction when the reference point reaches the processing end point determined by combining the duration of the voice obtained using the movement of power information and the temporal change in similarity. corresponding to the l-th (1≦l≦K) standard pattern for which K maximum similarities or minimum distances with K standard patterns are calculated, and the maximum or minimum value of the K similarities or distances is obtained. Then, add the l-th K values of the K x K bias tables prepared in advance and the K similarities or distances, and use the standard pattern to obtain the maximum or minimum value. A speech recognition method characterized by outputting corresponding speech as a recognition result. (2) The voice recognition method according to claim 1, wherein the presence/absence of voice is detected using a ratio of a voice signal to noise. (3) The speech recognition method according to claim 1, characterized in that the degree of similarity or distance between the characteristic parameters of the unknown input signal and the standard pattern of each speech is calculated using a statistical distance measure. (4) A patent claim characterized in that the statistical distance measure is any one of a measure based on posterior probability, a linear discriminant function, a quadratic discriminant function, Mahalanobis distance, Bayesian judgment, and a measure based on composite similarity. The speech recognition method described in scope 3.